Advancing Causal Analytics Using Biomedical Data

Abstract

Understanding the distinction between association and causality is crucial in the fields of health informatics and biomedicine, as causality allows for the modeling of manipulable relationships. Various statistical methods have been developed as alternatives to randomized clinical trials, which often are impractical due to ethical or cost considerations. The process of causal analysis typically involves two steps: causal structure discovery and causal effect estimation (also known as causal inference). The process of extracting causal structure from observational data, known as computational causal structure discovery (CSD), is an emerging field that has garnered considerable attention in recent years. Once the causal structure is known, or partially understood, the estimation of specific causal effects can be undertaken using causal inference (CI) methods. As vast biomedical data repositories continue to emerge, understanding how to effectively process causal structure discovery and causal effect estimation has become crucial for better utilization of observational biomedical data. This thesis aims to contribute by exploring existing causal discovery methods and developing new methodologies to address practical challenges in discovering causal relationships from biomedical datasets. The work is composed of four chapters that explore the use of existing methodologies to discover causal relationships among biomarkers related to Alzheimer's disease (AD), estimate causal effects related to Alzheimer's disease, and propose two methods to address data challenges inherent in electronic health records (EHRs) datasets. Ultimately, the research presented in this thesis offers practical examples of applying CSD and causal inference methods to address biomedical problems. It also proposes two novel methods to navigate prevalent data challenges, which are crucial for effectively utilizing EHR data and extracting meaningful causal relationships.